[BUG] Condensing context fails with Error 400 on Local Openai compatible endpoint

[huntyh]

Describe the bug

Sometimes when using Zoo code with a local openai provider (SGLang endpoint with a local model), prompt condensing will fail with 400 Bad request on the chat completions endpoint.

To Reproduce
Steps to reproduce the behavior:

1. Open a workspace
2. Open Zoo code, send a request that requires the AI to read a large amount of code
3. Context fills up, condensing context beings
4. Error 400 appears
5. The Agent can't do the task because the chat history will be continuously truncated

Expected behavior
Condensing context runs successfully

Screenshots

What version of zoo are you running

Version: 3.62.0 (40660f1)

Additional context

Logs from SGLang only indicate a bad request:

[2026-06-25 10:10:33] Decode batch, #running-req: 1, #full token: 51338, full token usage: 0.66, mamba num: 2, mamba usage: 0.07, cuda graph: True, gen throughput (token/s): 35.49, #queue-req: 0

[2026-06-25 10:10:34] INFO:     10.1.0.105:50783 - "POST /v1/chat/completions HTTP/1.1" 400 Bad Request

Note: this started like a week ago, before then it worked perfectly and I did not change anything on the sglang side.

Note 2: When this happens, starting a new chat (task) with the exact same prompt can fix the issue, and condensing runs as normal.